A quality score summarises specified checks on a specified population; it is not proof that a dataset is trustworthy for every use. Publish its components and denominators, keep critical exceptions visible, and prevent a high average from overruling a limitation that invalidates the intended decision.
Define precisely what earns a passing result
Write each component as a test with an eligible population. Date validity might mean parseable dates among records requiring a date. Source completeness might mean expected deliveries received by a deadline. These denominators differ and should not be casually pooled. Explain inapplicable fields and records that could not be tested. Excluding every difficult case from eligibility can raise the score while reducing useful coverage.
Distinguish validity, completeness, uniqueness, consistency and verified accuracy where they are measured. A populated field can still be wrong. A plausible date can belong to the wrong transaction. If no source-based accuracy check occurred, do not label a collection of formatting checks an accuracy score. Use component names that reflect the evidence actually produced and retain the date and population to which each result applies.
Choose aggregation rules deliberately
An unweighted mean gives a minor formatting check the same influence as a missing business channel. A field-count average can favour datasets with many easy columns. If a weighted score is useful, explain why the weights represent the intended use and show the component values. Do not imply that the arithmetic establishes a scientifically unique ranking; alternative weights can produce different priorities.
Add explicit publication gates for critical conditions. For example, a comparison of all channels may require every material source to be present, irrespective of the overall score. A gate should state the consequence and exception authority. Keep warning text beside the score so a reader does not need to open a methodology document to discover that a green indicator excludes the main source of uncertainty.
Use the score to support investigation
Track a stable set of checks over time and annotate changed definitions. A score can fall because new tests reveal old problems, or rise because difficult checks were removed. Neither movement necessarily reflects an actual change in the data. Preserve component history and enough detail to reconstruct the aggregate. Inspect concentrations by source and period so a strong global score does not hide a weak operational segment.
The government framework supplies background on quality dimensions; it does not prescribe the invented weights below. Treat the summary as an entry point to evidence and action. Pair unresolved exceptions with owners and review dates, and test whether the intended reader can answer what remains unassessed. A score without that explanation can encourage certainty that the underlying checks never established.
Document whether the score describes records, fields or source deliveries in its visible label. Readers should not have to infer that denominator from a percentage symbol or a colour.
Fictional dataset: 98% passes, a channel is still missing
A fictional scenario to explain the method, not a customer result or a market benchmark.
A fictional reporting team runs ten equally weighted formatting checks, each scoring 98%, and displays an overall 98%. Those checks apply only to the 5,000 rows that arrived. A separate expected-source roster contains five channels, but only four delivered data. Source coverage is therefore four of five deliveries, or 80%, with the missing channel's current value unknown.
The analyst keeps the formatting score but labels it as such. The all-channel sales comparison fails a publication gate because the absent channel is material to that decision. A direct-channel-only view can still be shown if its scope is explicit and its own checks support that use.
The report does not average 98% and 80% into a new reassuring number. The fictional case illustrates why test performance and population coverage require separate explanations. It does not demonstrate a universal scoring model or establish that the received rows are accurate.
TAKE THIS INTO YOUR NEXT CONVERSATION
A practical next-step checklist.
- Name each dimension and eligible denominator.
- Disclose weights and untested populations.
- Keep critical publication gates separate.
- Version checks and inspect source-level concentrations.
Common questions.
Is 98% quality enough to publish?
Only the intended use and remaining exceptions can answer that. A small unresolved issue may invalidate a critical decision.
Should all dimensions be averaged equally?
That is a policy choice, not a default truth. Show components and justify any weighting against the report’s use.
Further reading & source context.
These primary references provide background. They are not evidence for the fictional example above and do not endorse Accuralytix.
Found an error or a missing qualification? Send an editorial correction.
Check my data readiness.
Use a practical tool to turn this idea into a clearer starting point.
Check my data readiness